Do AI code reviewers actually catch real bugs?

We've been running an AI reviewer on every PR for a couple of months. It's great at catching missing null checks and inconsistent naming, but it consistently misses the stuff that actually bites us: race conditions, transaction boundaries, and logic that is wrong but plausible. The result is a wall of low-value comments that people start ignoring, which is arguably worse than no review at all. Curious how others are tuning this. Are you running it as a blocking gate or advisory only? Any prompt or